AQEA Gen-5 · Substrate Properties
The six load-bearing properties of the AQEA substrate.
Why Gen-5 is a generation distinction, not an incremental improvement. P1–P6 explained for engineers and technical reviewers.
P1
Multi-channel Orthogonal
The substrate is decomposed into a finite, fixed set of channels. Operations on different channel groups are mutually non-interfering — encoding information into one channel does not perturb the content of another. Empirically validated with measurably zero cross-channel interference. Float vectors from any encoder do not have this property: every coordinate participates in every operation.
Empirical Anchor
Δ = 0.00 pp (exact spectral-orthogonality)
P2
Byte-Deterministic Cross-Platform
The encoder produces byte-identical output across ARM NEON, x86 AVX-512, NVIDIA Vulkan and Apple Metal — verified across nine independent test fixtures. A customer encodes a corpus once and ships the encoded artefact to any platform without re-validation. This is a substrate-and-encoder property, not a property of the search algorithm running on top of it.
Empirical Anchor
26 / 26 tests PASS cross-platform · 9 / 9 hash-verification fixtures
P3
Structural Compression with Ranking Preservation
The encoded representation is approximately 23–29× smaller per document than the float input, and distances computed in the substrate rank documents in the same order as the underlying float-space — bit-identically at the Top-K level when the float reference is itself exact. The compression is structural (information reorganised into channel decomposition), not lossy quantisation.
Empirical Anchor
23–29× compression with bit-identical Top-K
P4
Encoder-Family-Agnostic
The ranking-preservation property holds across both transformer-learned encoders (BGE for text, WavLM for speech, ESM-2 for protein — three independent transformer families) and classical hand-engineered encoders (FFT-spectral signal-processing pipelines). P4 distinguishes AQEA from compression schemes that exploit transformer-specific output statistics.
Empirical Anchor
13 domains × 6 transformer-encoder families + 4 classical-DSP, all above 80% floor
P5
Noise-Resistant Ranking on Signal-Domains
On encoder families that produce noise-bearing float-output — specifically classical signal-processing encoders over sensor streams — distances in the substrate can exceed the float-baseline's nearest-neighbour quality, not merely preserve it. The mechanism: the float-mantissa contains high-frequency variance from sensor noise; trit-discretisation collapses this noise-floor, leaving distance computation to rank by signal-relevant differences.
Empirical Anchor
voraus-ad 133% · Digit_Fall 174% R-Ratio (Supra-Trit)
Phase-J · Falsification Box
Mass-Spec: 96.7% PASS-tier, not EXCEEDS.
On mass-spectrometry, AQEA achieves 96.7% retrieval-equivalent — a PASS-tier result, but it does not exceed the float baseline as it does on robotic-arm and wearable-fall signals. The reason is in the data: archive pre-processing strips per-bin noise from the mass-spec stream, so there is no high-frequency variance left for trit-discretisation to collapse.
We publish this as a feature, not a weakness. P5's mechanism is well-specified enough that we can name in advance which signal-domains will benefit and which will only match.
P6
Reversibly Decodable into Pareto-Front
A trainable inverse-decoder reconstructs a float-representation from the substrate, with task-preservation measured against the substrate's own direct-ranking baseline. Mode-selection is a deployment-time choice — the same substrate-encoded artefact is decoded in any mode without re-encoding the corpus.
Empirical Anchor
3-mode Pareto-front + task-elevating regime at msmarco-100k
Audit Mode
≥ 96.7%
per-vector cosine ≥ 0.90 · audit-trail workloads
General Mode
≥ 98.7%
per-vector cosine ≈ 0.75 · balanced fidelity / discrimination
Pure Retrieval Mode
≥ 99.7%
per-vector cosine ≈ 0.65 · msmarco-1M
Task-Elevating Regime
101.15%
at msmarco-100k — decoded ranking exceeds direct-substrate ranking
Closing
Why this is a generation distinction.
(P1) + (P2) + (P3) define the static substrate. (P4) + (P5) extend it across encoder paradigms and noise-bearing pipelines. (P6) gives the substrate a reversibly-decodable companion that downstream applications use without re-encoding.
A Gen-4 embedder produces an unstructured float vector and inherits its lack of internal channels. AQEA is not a post-processing step on that vector. It is a different output type, produced by an encoder whose codomain has structure that the Gen-4 codomain does not. Multi-channel co-encoding, deterministic content-addressing across hardware vendors, structural compression with bit-identical ranking — all follow from the substrate's structure, not from any single algorithmic trick.
The construction of the substrate, the channel decomposition, the deterministic encoder pipeline and the trainable inverse-decoder are patent-pending and not disclosed in this page or the public whitepaper. The intended integration surface for partner engineering teams is a black-box SDK with encoder and decoder behind a stable API.
Engineering-deep partner conversation?
P1–P6 are pre-registered. The Phase-J record is public. We respond within one business day.